{"id":"85d4442d-a82f-4525-88b4-d92b9645e65e","arxiv_id":"2412.08650","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A PINN trained on sparse FEM data plus Poisson's equation predicts electric field and potential for a capacitive sensor at finger distances from 0 to 25 mm with low overall error.","lead":"The paper trains a physics-informed neural network to predict electric fields and potentials around a capacitive touch sensor, replacing slow finite-element simulations with near-instant inference. The value is a fast surrogate that could accelerate automotive and smartphone sensor design.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's key generalization evidence may be circular: the 26 'unseen' test distances appear to be the same 26 distances used for PDE-loss training, so the reported NRMSE does not demonstrate prediction for any genuinely new finger distance or experimental setup.","rationale":"I read the paper in good faith as a proof-of-concept that a PINN with sparse FEM data and physics losses can reproduce electrostatic fields for a fixed capacitive sensor geometry as a function of finger distance. The reader identified the fixed geometry as the weakest assumption, and I agree that this limits the 'different experimental setups' claim. However, I see a more immediate, concrete threat to the evidence itself: the paper never states that the 26 evaluation distances are distinct from the 26 distances used in the PDE-loss training. The symmetry of the numbers, combined with the absence of any train/test split description, makes it plausible that the model was evaluated at distances it had already seen during training. Since finger distance is the only non-spatial variable, this would render the headline accuracy numbers uninformative about generalization to new setups. I do not claim the authors were deceptive; the text is simply ambiguous, and the most natural reading is that the same 26 distances were reused. The proposed concrete test would settle this directly by holding out distances. If the concern lands, the paper should be revised to either (a) demonstrate comparable accuracy on genuinely unseen distances, or (b) narrow the central claim to interpolation over a discrete set of trained distances. The reader's CONDITIONAL verdict already captures the need for such clarification, so I recommend no change to the verdict level, though the condition should explicitly include a distance-based holdout evaluation.","tokens_in":6899,"tokens_out":7486,"duration_ms":71223,"concrete_test":"Inspect the training loop's distance sampler. If it draws from a discrete set such as {0,1,...,25} mm, then retrain with only even-mm distances in the physics loss (D_train = {0,2,4,...,24}) and evaluate at the odd-mm distances (D_test = {1,3,5,...,25}), using the same loss weights and collocation strategy. If the held-out NRMSE remains within the ranges in Table 1, the concern is resolved; if it degrades, the reported errors are an artifact of evaluating at trained distances. As a separate check for the 'different experimental setups' claim, run the trained network at a sensor width not present in the training data; without geometry inputs, this should fail, confirming that the surrogate is limited to the fixed sensor/PCB geometry.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the trained PINN predicts electrostatic quantities for 'different experimental setups' without running simulations, with evidence from 'unseen test cases.' The only setup parameter varied in this work is finger distance d, and the evaluation section is ambiguous about whether the test distances are truly unseen. In 'Experiments and Results,' the paper states: 'PDE-based loss minimization during training is conducted with 26 different finger distances to the sensor, ranging from dmin = 0mm to dmax = 25mm.' In the Conclusions, it says: 'we generated experimental data for 26 different finger positions to the sensor and organized them into three proximity zones.' The identical count suggests the same 26 distances were used in both training (for the physics loss via collocation points and translated finger boundaries) and evaluation. If so, every test distance was already seen during training, and the reported NRMSE merely measures interpolation among trained values, not generalization to new distances. The paper does not describe any train/test split by distance, nor does it evaluate at distances outside the training set. This matters because finger distance is the only continuous non-spatial input; if the distance distribution is not held out, there is no empirical basis for the claim that inference can be carried out 'for different experimental setups without the need to run simulations.' Additionally, the sensor and PCB geometry are fixed (as the paper states: 'Sensor and PCB have fixed dimensions'), so any genuine change in electrode layout, finger shape, or environment is outside the model's input space. The most load-bearing weakness is therefore that the empirical core of the generalization claim rests on distances that may not be unseen at all.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a physics-informed neural network (PINN) surrogate for a capacitive touch sensor. The model takes spatial coordinates and finger distance as inputs and outputs the electric field E, electric potential V, and charge density ρ. The loss is a weighted combination of boundary-condition, physics-residual, and sparse FEM data terms; the physics residuals come from the electroquasistatic form of Maxwell's equations (Faraday's law, Gauss's law, the E–V gradient relation, and Poisson's equation). The network is trained on FEM data at three finger distances together with PDE collocation points sampled over 26 finger distances, and evaluated on what the paper calls an unseen dataset. Reported overall NRMSE values range from about 0.003 to 0.062 for E and V, with peak 25% NRMSE values of 0.82–0.88 near the finger boundary. The paper claims a 10x inference speedup over FEM and concludes that the PINN can serve as a surrogate for fast design optimization.","tokens_in":7203,"tokens_out":3242,"duration_ms":30392,"significance":"If the central claim is supported, the paper would contribute to an industrially relevant application of PINNs: accelerating capacitive touch sensor design. The approach is reasonable in that the physics loss is derived from Maxwell's equations rather than from the fitted model, and the combination of sparse FEM data with PDE residuals is a sensible way to regularize the solution. The reported errors on the evaluated distances are encouraging, and the demonstrated ability to predict fields in a 3D geometry with a finger at various distances is a useful proof of concept. However, the significance is substantially limited by the lack of a genuine held-out test set by distance, the fixed sensor/PCB geometry, the absence of code and data, and the lack of comparison to simpler baselines. These gaps, if not addressed, reduce the paper to a demonstration of interpolation over an input parameter that was already present in training, which is far weaker than the claimed generalization to 'different experimental setups.'","major_comments":[{"comment":"The claim of evaluation on 'unseen test cases' is not substantiated, and the paper appears to have a train/test circularity with respect to finger distance. The Experiments and Results section states that 'PDE-based loss minimization during training is conducted with 26 different finger distances to the sensor, ranging from dmin = 0mm to dmax = 25mm,' and the Conclusions state that 'we generated experimental data for 26 different finger positions to the sensor and organized them into three proximity zones.' These are the same 26 distances, with no description of any train/test split by distance. Because finger distance is the only non-spatial input that is varied, all evaluated distances may already have been used during physics-loss training, so the reported NRMSE values measure interpolation over already-seen distances rather than generalization to new distances. To support the central claim, the authors must re-evaluate on distances that are excluded from all training losses (including the PDE loss), or explicitly report a train/test split and show that the test distances were never used in any loss term.","section":"Experiments and Results / Conclusions"},{"comment":"The input space is only (x, y, z, d), and the manuscript explicitly states 'Sensor and PCB have fixed dimensions; thus, BCs are randomly sampled from their corresponding meshes during training.' The abstract and conclusion claim that the model can perform inference 'for different experimental setups,' but the only setup parameter that can vary in the current model is the finger distance for one fixed sensor and PCB geometry. Changes in electrode layout, sensor size, finger shape, or nearby environment cannot be represented. The claim in the abstract and conclusions is therefore overbroad; it should be restricted to 'different finger distances for a fixed sensor/PCB geometry' unless the model is shown to generalize across geometry variations, which it currently cannot by construction.","section":"Neural Network Setup"},{"comment":"The physics loss in Eq. (12) contains redundant or trivial terms. The term ||Ê + ∇V̂||^2 enforces E = -∇V. The second term, ||-∇ × (ε0εr∇V̂)||^2, is the curl of a gradient times a scalar when εr is constant; since εr is set to 1 throughout the experiments, this term is identically zero. The third and fourth terms both contain ∇·(ε0εrÊ), and the fourth term is a residual between ρ̂ and the divergence. The authors should clarify on which subdomains each residual is applied (e.g., Laplace in free space, Poisson on conductors), remove the redundant curl term or justify its presence with a spatially varying εr, and explain the role of each term. As written, part of the 'physics' in the loss imposes no constraint.","section":"Loss Functions, Eq. (12)"},{"comment":"The peak 25% NRMSE values for E components in the near zone are 0.819–0.880, which means that in the region closest to the finger the predictions are essentially unreliable, and this is precisely the region that matters most for touch detection. The paper attributes this to sharp transitions around the finger boundary, but does not provide error maps, the fraction of points affected, or a distance-from-boundary analysis. Without such information, the practical value of the surrogate for design decisions in the near zone remains unclear. The authors should quantify how localized these high errors are and, if possible, report errors on derived quantities such as capacitance, which is the most relevant output for capacitive sensor functionality.","section":"Table 1"}],"minor_comments":[{"comment":"There is a typo: 'gradiant' should be 'gradient' in the sentence 'Additionally, the gradiant and the curl of E are calculated...'.","section":"Methodology, Governing Equations"},{"comment":"The paper does not provide code, data, training details such as exact hyperparameter values, or the procedure for generating the 'unseen' dataset. For reproducibility and for the reader to assess whether the reported results are robust, the authors should make the data and code available, or at least describe the exact test data generation procedure and the train/test split.","section":"General"},{"comment":"The paper does not compare the PINN against a simpler baseline, such as a fully data-driven neural network trained on the same FEM data, or against FEM interpolation. A baseline comparison would help assess whether the physics loss actually provides the claimed benefit in accuracy and generalization, rather than merely acting as a regularizer.","section":"Experiments and Results"},{"comment":"The statement that 'PINNs maintain their fast inference capabilities across different resolutions of input' is not demonstrated anywhere in the paper; no resolution study is reported. Please either add supporting experiments or remove the claim.","section":"Conclusions"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant industrial problem and the use of a combinatorial PINN loss with sparse FEM data is reasonable. However, the evaluation appears to be circular with respect to the finger distance parameter, which is the central variable of the claimed generalization. The authors should be asked to re-run the evaluation on genuinely held-out distances, to restrict the claims to the fixed-geometry setup, and to strengthen the quantitative evaluation with error maps or derived capacitance errors. If the held-out evaluation still shows good NRMSE, a revised version could become acceptable; as it stands, the evidence does not support the stated conclusions. The lack of code and data also makes it impossible to verify the reported numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent engineering application of PINNs to 3D electroquasistatic sensor modeling, and it probably works as a proof of concept. But the paper's central claim—that the model predicts for 'different experimental setups'—is not supported by the evaluation as described. The same 26 finger distances appear in both the PDE-loss training and the test set, so the reported NRMSE is interpolation over already-seen distances, not generalization to new ones.\n\nWhat's actually new: applying a PINN to a 3D capacitive sensor with finger distance as an additional input, using sparse FEM data at three distances plus physics residuals. That's a legitimate extension of MaxwellNet (2D optics) and Zhang et al. (1D time-domain). The architecture and loss are standard—FNN, tanh, Adam, weighted data/physics/boundary losses—but the domain is new and the distance-conditioned input is a sensible way to parameterize the setup. The reported overall NRMSE of 0.003–0.062 for E and V is genuinely good if those numbers hold.\n\nThe soft spots are real and in proportion. First, the circularity concern: the paper says PDE-based loss minimization was conducted with 26 finger distances from 0 to 25mm, and then says they generated experimental data for 26 finger positions for testing. No train/test split by distance is described. Unless some distances were held out from the physics loss, the 'unseen test cases' aren't unseen. This directly undermines the generalization claim. Second, no code or data is released, so nothing is independently reproducible. Third, there's no baseline—a pure data-driven NN or a simpler interpolator—so we don't know whether the physics loss is actually buying anything. Fourth, the sensor and PCB geometry are fixed, so the model can't handle other electrode layouts or finger shapes without retraining; the abstract's 'different experimental setups' oversells that. The peak 25% errors (NRMSE ~0.82–0.88 at the finger boundary) are acknowledged, and they're acceptable for a proof of concept, but they'd matter in real use.\n\nWho this is for: engineers in automotive HMI or capacitive sensing who want a fast surrogate for design sweeps over finger distance on a fixed sensor geometry. It's a useful pointer, not a turnkey tool. The paper deserves peer review—the idea is sound and the execution is mostly careful—but it needs a proper held-out distance evaluation, baselines, and reproducibility artifacts before the claims can be trusted.","headline":"A reasonable 3D PINN surrogate for capacitive sensors, but the generalization claim rests on test distances that were already used in training, so the proof of generalization is weaker than it looks.","tokens_in":7792,"tokens_out":2381,"would_cite":false,"duration_ms":19346,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A physics-informed neural network predicts electric field and potential for capacitive touch sensors across 0–25 mm finger distances with a few percent normalized error and roughly 10x faster inference than FEM.","keywords":["physics-informed neural network","capacitive touch sensor","electrostatic simulation","surrogate model","Maxwell's equations","electric field prediction","finite element method"],"falsifier":"Generate a finite-element field solution for a capacitive sensor with a different electrode layout or a non-cylindrical finger, feed it to the trained network at a finger distance inside 0–25 mm, and compare predicted $E$ and $V$ with the simulation; if the normalized error jumps well above the reported 0.003–0.062 range, the surrogate does not generalize to changed experimental setups.","tokens_in":6707,"feed_emoji":"⚡","tokens_out":6885,"duration_ms":58625,"temperature":0.7,"pith_summary":"The paper tries to establish that a physics-informed neural network can serve as a fast surrogate for finite-element electrostatic simulation of a capacitive touch sensor. The network takes spatial coordinates and finger distance as inputs and outputs electric field, electric potential, and charge density, trained with a loss that combines sparse simulation data at only three finger positions with residuals of the electroquasistatic Maxwell equations. If correct, designers could query electric field and potential for any finger distance between 0 and 25 mm in seconds instead of running multi-physics simulations, with normalized errors mostly between 0.003 and 0.062 and roughly tenfold faster inference. This matters because capacitive touch sensors in automotive and consumer devices are tuned through expensive simulations.","feed_headline":"Physics-informed network models touch sensors 10x faster than FEM","feed_subtitle":"Design queries drop from long finite-element runs to seconds, with fields accurate to a few percent.","key_machinery":"The central object is the combinatorial loss function that couples data and physics. $L_{\\text{bc}}$ enforces Dirichlet boundary conditions $V=0$ on finger and PCB and $V=3.3$ V on the sensor; $L_{\\text{phy}}$ computes residuals of the electroquasistatic Maxwell equations using automatic differentiation with respect to coordinates, including $E=-\\nabla V$, $\\nabla\\times E=0$, and the Laplace/Poisson equations; $L_{\\text{data}}$ measures mean-square error against the sparse low-resolution simulation dataset $\\Omega_S$ at three fixed finger distances. Collocation points come from $\\Omega_S$ plus a non-uniform refined 3D grid $\\Omega_G$ denser near the sensor-finger setup, and finger distances for the grid points are sampled uniformly over $[0,25]$ mm. An adaptive loss-weighting algorithm balances the three loss terms during training. This combination lets the physics residuals regularize the network so it can interpolate to distances and locations not present in the sparse training data.","core_discovery":"In the paper's own terms, the contribution is a trained surrogate sensor model: a fully connected network with five hidden layers of 512 neurons and tanh activations that solves the electroquasistatic form of Maxwell's equations in a 3D domain. The inputs are $(x,y,z,d)$; outputs are $\\mathbf{E}=(E_x,E_y,E_z)$, $V$, and $\\rho$. The model is trained by minimizing $L_{\\text{total}}=\\lambda_1 L_{\\text{bc}}+\\lambda_2 L_{\\text{phy}}+\\lambda_3 L_{\\text{data}}$, where the physics losses enforce $\\nabla\\times\\mathbf{E}=0$, $E=-\\nabla V$, Laplace's equation in free space, and Poisson's equation at conductor surfaces, while the data loss anchors the solution to sparse simulation output at three finger distances. Evaluated on 26 unseen finger distances grouped into near, hover, and far zones, the network reaches NRMSE between 0.003 and 0.062 for $E$ and $V$, with the largest deviations in the $z$-component of the electric field near the finger boundary, and delivers roughly 10x faster inference than the finite-element simulation.","pith_inferences":["The fixed-geometry setting means the surrogate cannot represent a new electrode layout, sensor size, finger shape, or different dielectric environment without retraining; extending the input space to geometry parameters would be a natural next step beyond what is demonstrated.","Charge density $\\rho$ is used only inside the physics loss and is not reported as an output metric, so the paper's accuracy claims apply to $E$ and $V$; a reader should not infer that $\\rho$ is validated independently.","The roughly 10x inference speedup counts only inference time, not the cost of generating the sparse FEM dataset or the 800k-iteration training run; the practical gain depends on how many design queries amortize that fixed cost.","Since the model was trained with $\\varepsilon_r=1$ throughout the domain, wet-finger or moisture scenarios would need additional training or a permittivity input before the surrogate is useful in humid environments."],"forward_implications":["A design engineer can query electric field and potential for any finger distance in 0–25 mm in seconds, without regenerating a finite-element mesh or rerunning the simulation.","Sparse simulation data at only three finger distances, combined with PDE residuals, is enough to generalize across 26 tested distances, reducing the cost of building the training dataset.","Predictions remain accurate in near, hover, and far proximity zones, with normalized errors between 0.003 and 0.062 for $E$ and $V$; the main error concentration sits at the finger boundary.","Because the network is mesh-free and GPU-accelerable, the same trained model can be sampled at arbitrary resolutions and query points, unlike fixed-resolution finite-element output.","The framework is positioned as a foundation for modeling sensor arrays and long-term performance, since the per-query cost stays low when the geometry input set is extended."],"supporting_citations":[{"why":"Supplies the electroquasistatic approximation that lets the model drop magnetic-field terms and solve only electrostatic equations.","marker":"Baxter 1996"},{"why":"Establishes the PINN methodology of embedding PDE residuals in the neural network loss function.","marker":"Raissi et al. 2019"},{"why":"Introduces the combined data-driven and physics-driven loss structure the paper adapts for convergence speed.","marker":"Ma et al. 2020"},{"why":"Provides the adaptive loss-weighting algorithm used to balance boundary, physics, and data losses during training.","marker":"Heydari et al. 2019"},{"why":"Cited for PINN generalization difficulties near boundaries, which the paper invokes to explain larger errors at the finger edge.","marker":"Sharma et al. 2023"}],"fun_headline_variants":["Touch sensor design: PINN solves Maxwell's equations in seconds","Neural net solves Maxwell's equations for touch sensors","Physics-informed network speeds up touch sensor simulation","10x faster touch sensor modeling via physics-informed neural net","Maxwell's equations baked into neural net for faster touch sensors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that the trained model serves as a surrogate for different experimental setups rests on the premise that changing only the finger distance spans the relevant design space; the sensor and PCB geometry, finger shape, and material permittivity are fixed during training.","fun_headline_variants_meta":{"raw":{"variants":["Touch sensor design: PINN solves Maxwell's equations in seconds","Neural net solves Maxwell's equations for touch sensors","Physics-informed network speeds up touch sensor simulation","10x faster touch sensor modeling via physics-informed neural net","Maxwell's equations baked into neural net for faster touch sensors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00058,"raw_usage":{"total_tokens":2746,"prompt_tokens":972,"completion_tokens":1774,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":1695}},"tokens_in":588,"tokens_out":1774,"duration_ms":12991,"temperature":1.0,"reasoning_tokens":1695,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:11:20.023887+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a finite-element field solution for a capacitive sensor with a different electrode layout or a non-cylindrical finger, feed it to the trained network at a finger distance inside 0–25 mm, and compare predicted $E$ and $V$ with the simulation; if the normalized error jumps well above the reported 0.003–0.062 range, the surrogate does not generalize to changed experimental setups.","supporting_citations":[{"cited_title":"Capacitive Sensors: Design and Applications","cited_arxiv_id":null,"evidence_quote":"Supplies the electroquasistatic approximation that lets the model drop magnetic-field terms and solve only electrostatic equations."},{"cited_title":"Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations","cited_arxiv_id":null,"evidence_quote":"Establishes the PINN methodology of embedding PDE residuals in the neural network loss function."},{"cited_title":"Softadapt: Techniques for adaptive loss weighting of neural networks with multi-part loss functions","cited_arxiv_id":null,"evidence_quote":"Provides the adaptive loss-weighting algorithm used to balance boundary, physics, and data losses during training."},{"cited_title":"T., 2023","cited_arxiv_id":null,"evidence_quote":"Cited for PINN generalization difficulties near boundaries, which the paper invokes to explain larger errors at the finger edge."}],"review_version":1}